Qwen3.8-Flash-Next Gyro S on a Valve Steam Machine (8 GB AMD GPU, 32 GB RAM): about 18 t/s decode, about 235 t/s prefill, stock SteamOS

I followed up on the post about running Qwen 3.8 Flash Next with 8 GB VRAM and 32 GB RAM. I wanted to see how it does on my Steam Machine. Setup Valve Steam Machine, stock SteamOS 3.9.5, AMD Navi 33 with 8 GB, 32 GB RAM, NVMe Stock Strata terminal 0.2.9 (engine 0.1.55, HIP package), Gyro S TQ1_0 128K context, q4_0 KV, no vision, MTP off, served on the LAN Results (Strata's own timing line from the engine log, about 8K-token prompts, 3 runs each) Config Prefill t/s Decode t/s Default ( --tune skip ) 219-240 15.1-15.9 Quick tuner 237-239 14.5-17.2 (tuner said "already fastest") pcie-frac 0.35 239 14.8-16.0 kv-resident 20480 + vram-reserve-mib 900 227-229 14.8-15.5 resident-budget-gib 64 232-242 17.3-19.8 With the final config at 128K context, 58K and 61K-token prompts gave 234 and 236 t/s prefill (about 4 minutes each) and 17.6-17.9 t/s decode. No OOM, swap went from 0.2 to 0.3 GB, and available RAM bottomed out around 3.4 GB. I did not test beyond 61K tokens or with a busy desktop session. What mattered --set resident-budget-gib=64 was the only change that helped, about +20% decode. It makes the engine keep the experts in its own RAM tier. Without it they sat in the OS page cache, so my RAM monitor showed under 5 GB "used" while the machine held about 28 GB of cache. Prefill stayed at about 235-242 t/s in every config I tried. pcie-frac 0.35 , a larger kv-resident , and a bigger VRAM reserve did nothing measurable here. The OP's pcie-frac was calibrated for their own machine, and the quick tuner changed nothing either. MTP: I ran with --mtp off and tested real MTP to check. On my install --mtp on silently did nothing. The saved config said mtp: true , but the draft layer was never downloaded or passed to the engine ("+0 MiB for the draft head" in the log). I had to build it by hand ( mtp_fetch.py , mtp_pack.py , mtp_rt.py from the Strata source) and pass --set mtp= . With the draft layer really loaded (820 MiB of VRAM, expert cache cut from 2225 to 1179 slots), decode was 15.6 / 18.4 / 16.9 t/s, no gain over 17.2, and prefill was 210 vs 232 t/s, about 10% lower. On 8 GB of VRAM it doesn't pay off. That was 3 runs at 8K context only. I did not try --adapt-* , --spec-min-p , or the thorough tuner. SteamOS tips Anything started from an SSH session dies when the session ends. Use systemd-run --user --property=KillMode=process --property=RemainAfterExit=yes strata launch ... . --no-tui shows no download progress, so check the file size instead. Check the generated config under ~/.local/state/strata/launch/ to see which --set values were really applied. Unknown names only warn. Command for the final setup (I saved it as a preset with --save-preset ): strata launch --model gyro-s --family gyro --context 128k --kv q4_0 --vision off --mtp off \ --set resident-budget-gib=64 --access lan --port 8080 --no-tui --yes --restart Caveats Decode varies a lot between runs with speculative draft acceptance, so treat the numbers as ranges. With resident-budget-gib=64 only about 3.5 to 4 GB of RAM stays available. My machine rebooted once during testing with a desktop session logged in (cause not confirmed), so close other apps. Single user, one request at a time, 3 runs per config. I could not find a ready-made Strata config for 8 GB VRAM with 16-32 GB RAM. If you have one, please share it. All scripts are on the Github repo .